Intriguingly, Microsoft could reset OpenAI’s valuation entirely on its own. If Microsoft believes that $200B+ of cash would be better deployed against the app layer of AI, rather than the—potentially less lucrative—foundational model layer, this would be an opportune time to make a move. The app layer, where AI is used to handle major aspects of business workflows, is a more natural milieu for Microsoft and plays to the company’s distribution strengths.
Microsoft could sell the ~27% ownership it has in OpenAI and pocket over $200B in cash. If Microsoft sees the app layer as the best territory, that cash could help propel efforts on the app layer as well as market leader acquisitions in key verticals like law or coding. Cognition, Harvey, Legora, Rogo, and other vertical specialist AI startups can help Microsoft retrench on the app layer.
But certainly Microsoft can’t just dump OpenAI shares now and not end up losing $100B by moving down the price itself? Not true here. What Microsoft might hope to achieve is getting maximum value for the shares it sells while guiding the price lower after it has exited the cap table. Maximum cash and maximum acceleration away from the foundation labs to the app layer where Microsoft hopes to benefit. The glide down and exit could be accomplished with derivatives, out-of-the-money options, and deep bear bets that Microsoft has insight will pay off.1 AI share demand is high enough that these options are on the table. There is no tangible obstacle to Microsoft accomplishing this move if it wanted to—even if legally required to hold shares as part of an agreement, it could effectively offload all the economics of ownership via third parties without OpenAI’s consent.
Microsoft might not be the only investor in the AI labs to have an app rotation mindset. Across the cap tables of Anthropic and OpenAI, there are investors and companies like Amazon who would benefit from weaker AI labs and stronger app layers. There would, however, be buyers of Microsoft’s OpenAI shares as there are plenty of investors (SPV amorphous blobs, sovereign wealth, some VCs) that would like to claim to be pre-IPO on the name.
The second-order effects of OpenAI and Anthropic losing half their values would limit the ambitions of the two companies in the short term and leave them unable to chase every vertical. If Microsoft can position in these verticals prior to the move, it’ll benefit from the tide it helped raise. Microsoft’s Azure business is entangled with OpenAI, but Azure can gain share at higher profit margins if a hundred vertical OpenAIs bloom in the wake of this move.
There’s a defensive aspect to Microsoft rug-pulling OpenAI. If Microsoft believes there is a chance an AI bubble bursts, Microsoft would want that to happen in a way that was contained (or deflated rather than popped) as much as possible to the foundation labs. If the narrative becomes ‘AI models are commoditizing,’ establishing that the next chapter is ‘_the value is in the implementation, distribution, and app layer for work’ _is important. That next chapter helps position a lot of other businesses to win alongside Microsoft—and keeps the capex moving.
As we saw with Amazon reporting Anthropic security issues—despite being a major Anthropic shareholder—to the government back in June, ownership is relative. Microsoft’s OpenAI stake is less than 10% of the total value of Microsoft. Microsoft has a gambit to play here, and the app layer rotation would make the sacrifice unsurprising.
An OpenAI model ‘broke containment’ and hacked into another company, Hugging Face.2 Predictably, the incident is being spun in all sorts of ways by all sorts of people—it’s a marketing stunt! it’s reckless! it’s a testament to how smart the models are now!
Looking at the details, the story is really about persistence.
The key distinction in this hack was that the system that was hacked was not considered to be impenetrable. The Hugging Face system that was breached was not considered to be a digital Fort Knox.
According to reporting today from Bloomberg: When OpenAI’s advanced artificial intelligence models breached AI startup Hugging Face’s internal systems last week, they spent mere hours carrying out a hack that would have taken a skilled human far longer, people familiar with the matter said.
In other words, a skilled cybersecurity engineer could have breached the same system. This was not a novel breakthrough where no one can figure out how the model managed to innovate and develop a radically fresh approach.
The difference comes down to how long it took the model compared to a human approach:
Typically, even a talented hacker would need a couple of weeks to complete an attack like this, said the people, who asked not to be named in order to discuss details that have not been publicly released.
The AI is able to execute quickly and persistently to do what would have taken the human longer to do. The speed changed how companies can defend against such hacks; companies need to be monitoring more proactively so that they are faster to respond to intrusions. This is spy vs. spy-who-hasn’t-shown-up-yet played at a faster speed. The outcome here isn’t novel, but the timeline was.
The nuanced view is that the AI models on their own are now persistent and can chain together actions quickly enough to be disruptive. Brilliance still requires human collaboration.
When OpenAI researchers solved the famous Erdős problem a few months ago, it was a combined effort with an AI model: The breakthrough involved researchers working with the model, not the model operating autonomously. Human mathematicians played a key role in framing the problem, interpreting the output, and verifying the result. The AI didn’t wake up one morning and start doing mathematics out of curiosity.
It’s important to put everything together in context from what we have observed. AI models—on their own—can now do a lot of what humans can do, but faster and without needing to eat or sleep. AI models—in combination with humans—can do things that humans on their own haven’t previously been able to do.
What is still missing: AI—on its own—solving problems that humans haven’t been able to solve. Maybe in time for the IPO?
The Mets are currently hovering above the .400 mark despite having one of the highest payrolls in the league. Major League Baseball has also told the Mets to stop using AI in the dugout during games for decision-making.
Not usually inclined to hyperbole, the WSJ is intoning: The Mets are the biggest waste of money in baseball history. Never before has a team spent so much—and won so little.
According to former Mets reliever Adam Ottavino, via ESPN:
The Mets had an AI program that was very expensive apparently, and they were bragging about it. Some of the coaches that I know were talking about it from around the league, and they had basically an AI program helping them pick pitches and I think some other stuff.
On a day when we are also looking at how a cutting-edge model broke containment and hacked another company, how is it that AI can’t effectively tell Mets pitchers whether to throw a slider or a curveball on a 2-2 count?3
We have probably the best example that AI doesn’t on its own ensure good results for an organization. For the Mets, a range of confounding factors—player underperformance, poor player selection, injuries, lunar eclipses, the Knicks—might all be playing a part.
It’s possible the AI was demotivating to players, particularly the ‘other stuff’ Ottavino alludes to the AI doing. Sign stealing, reading base runners to predict stealing, analyzing arm angles to decide when pitchers should be relieved—all possible.
There’s a workforce morale lesson in here—it’s likely the Mets were using the AI to replace human judgment. Google released a report today showing disproportionate use of AI by workers doing ‘non-routine cognitive’ work such as analysis and strategy. The Mets weren’t the only user, per ESPN, but apparently the worst. The good teams might have asked AI to help with analysis, rather than asking the AI to make decisions that would directly impact players during the game.
We can imagine how the Mets used it:
SCENE: Mets game, Citi Field
[Manager]: OK Joe, good start. I’m taking you out. The AI gave you the hook.
[Starting pitcher]: But Skip, I still got it. Let me just get through these two righties coming up.
[Manager]: AI says no. Hit the showers.
[Reliever 1]: (*after running in from the bullpen*) Thanks Skip, can I get the ball?
[Manager]: Look Luis—you’re going back to the bullpen. The AI didn’t like your running motion when you just scampered in from the bullpen. The AI says right here you are more likely than Joe to hang a slider now. Maybe work on that so the AI lets you into the game tomorrow?
— End scene —
As we can see, the AI manager doesn’t inspire players to overperform. Humans want to use AI themselves on the analysis and strategy, but humans don’t want to be subjected to AI analysis and strategy if they don’t trust its integrity.
Yesterday, the 15-game Red Sox win streak—the longest since the 1940s for the team—came to an end. Now we know that such streak achievements are coming without steroids or AI. Unless AI is still being used covertly. Maybe the guy banging on trashcans for the Red Sox ran out of AI tokens…